Skip to content

Latest commit

 

History

History
107 lines (99 loc) · 5.28 KB

README.md

File metadata and controls

107 lines (99 loc) · 5.28 KB
Raw

Installation

Create a ROS catkin workspace with the following packages:

Documentation

Gripper Operation

  • Start soft gripper: roslaunch softgrasp_ros hand_control_interface.launch
  • Grasp params: rosservice call /grasp_params 0.5 0.5
  • Grasp close: rosservice call /grasp_state true, grasp open: rosservice call /grasp_state false

Moving to start pose

roslaunch softgrasp_ros move_to_start_pose.launch n:=0, replace 0 with pose number

Coordinate frames:

  • World = robot base. Orientation: X facing front of the robot, Z up, Y left
  • Stiffness frame:
    • Position at the contact point between plate and fingers
    • Orientation: X perpendicular to finger closing motion facing away from the robot, Z down, Y right
    • Sim: env_eef_{xpos,xquat}, Real: RobotState.O_T_EE (RobotState.EE_T_K is identity)

Softgrasp TF Policy

  • Launch TF policy on real robot:

    • Specify policy and checkpoint in softgrasp_ros_controllers/real.launch
    • Launch policy: roslaunch softgrasp_ros_controllers real.launch policy_name:=softgrasp. It starts a TF softgrasp policy action server
    • You can change policy_name to zero for straight line to goal
    • Switch from cartesian_impedance_example_controller to tf_policy_controller using
    $ rosservice call /controller_manager/switch_controller "start_controllers: ['tf_policy_controller']                                                       
stop_controllers: ['cartesian_impedance_example_controller']
strictness: 0
start_asap: false
timeout: 0.0"

  • In above launch commands, replace real.launch with sim.launch to launch the same policy in Gazebo

    Testing the TF policy in Gazebo

    • roslaunch softgrasp_ros_controllers sim.launch policy_name:=softgrasp
    • Teleoperate through RViz to grasp the stone and lift it up
    • Close the grasp by rostopic pub --once /franka_gripper/grasp/goal franka_gripper/GraspActionGoal "goal: { width: 0.03, epsilon:{ inner: 0.005, outer: 0.005 }, speed: 0.1, force: 5.0}"
    • Open the grasp by rostopic pub --once /franka_gripper/move/goal franka_gripper/MoveActionGoal "goal: { width: 0.08, speed: 0.1 }"
    • Switch from cartesian_impedance_example_controller to tf_policy_controller using commands in previous section

Experiments

  • Move to a pre-sampled random start pose: roslaunch softgrasp_ros move_to_start_pose.launch n:=6 (optional)
  • Move to goal vicinity: roslaunch softgrasp_ros run_experiment.launch policy_name:=<softgrasp|downward|random_search|zero>. This also starts a TF downward policy action server
  • Switch from position_joint_trajectory_controller to tf_policy_controller (note, command is different from above, which stops the cartesian_impedance_example_controller):
$ rosservice call /controller_manager/switch_controller "start_controllers: ['tf_policy_controller']                                                       
stop_controllers: ['position_joint_trajectory_controller']
strictness: 0
start_asap: false
timeout: 0.0"
  • In the rqt_controller_manager window, stop the running position_joint_trajectory_controller and start the downward_force_controller

Policies used

  • Translational stiffness = rotational stiffness = 1200
  • translational damping = rotational damping = 4 * sqrt(1200)
  • translational action scaling +- 0.15
  • translational action scaling +- 0.5
  • force scaling = 15.0
  • torque scaling = 7.5
  • ours (full), ours w/ medium and small plate: exp_137 @ 52K (i.e. policy_checkpoint_0005200000)
  • ours w/o delay model exp_147 @ 64K
  • ours w/o history exp_157 @ 64K
  • ours w/o noise exp_148 @ 64K
  • ours w/ 16 steps history exp_158 @ 64K
  • recurrent exp_177 @ policy_checkpoint_0002025000
  • with velocity observations exp_181 @ policy_checkpoint_0006400000
  • PPO: policy from Ankur, TF action server g_T_ee['position']['x'] + 0.07
  • downward with medium plate: vicinity_height_offset = 0.1, O_T_K_goal.position.x += 0.05 in move_to_goal_vicinity.py
  • Lee et al (16, 16): exp_173 @ 64K
    • for medium plate, TF action server g_T_ee['position']['z'] - 0.02
    • for medium plate, TF action server g_T_ee['position']['z'] - 0.04
  • softgrasp_ros_controllers/config/softgrasp_ros_controllers.yaml success_thresh_z:
    • large (main) plate -0.02
    • medium plate -0.04
    • small plate -0.04
  • Also for the small plate experiment:
    • run_experiment.launch: vicinity_height_offset = 0.07 and use static_poses_far.launch
    • translational stiffness = 2400, rotational space = 1700
    • TF action server g_T_ee['position']['z'] - 0.04
  • ours cup
    • g_T_ee['position']['x'] + 0.09
    • g_T_ee['position']['y'] - 0.05
    • g_T_ee['position']['z'] - 0.15
    • success_thresh_z -0.08
    • gains: 2400, 1700
  • lee et al (16, 16) cup
    • g_T_ee['position']['x']
    • g_T_ee['position']['y'] + 0.01
    • g_T_ee['position']['z'] - 0.2
    • success_thresh_z -0.08
    • gains: 2400, 1700
  • random cup
    • move to goal vicinity O_T_K_goal.position.x += 0.05
    • force_limit 30
    • g_T_ee['position']['x'] + 0.05
    • g_T_ee['position']['y']
    • g_T_ee['position']['z'] - 0.2